Natural Yet Challenging to Detect: Robust In-the-Wild TTS through EMA and Dual-Scoring Prompt Selection -- Submission for WildSpoof 2026 TTS Track

Abstract: In this technical report, we describe our submission for the WildSpoof Challenge TTS Track: Text-to-Speech with In-the-Wild Data. We introduce F5-TTS-DPS, a model built upon the F5-TTS architecture. Our approach integrates Exponential Moving Average (EMA) into supervised fine-tuning to stabilize training and improve generalization. To enhance synthesis fidelity, we leverage LLMs and large audio LLMs (LALMs) for dual-scoring prompt selection, filtering reference audio and text prompts to ensure quality while addressing alignment issues in noisy datasets. Experimental evaluation demonstrates that F5-TTS-DPS achieves strong performance with UTMOS of 3.20 and speaker similarity of 0.51 on the development set. More importantly, our model achieves the best a-DCF scores of 0.1582, 0.5233, and 0.2562 across three advanced SASV systems among all submissions, indicating our synthesized speech is the most difficult to detect and exhibits the highest degree of naturalness and authenticity. Combined with competitive WER performance, these results validate the effectiveness of our approach in generating natural-sounding speech with strong spoofing capabilities.

• The paper demonstrates that integrating EMA fine-tuning stabilizes training and improves robustness against noisy, real-world datasets.
• It introduces a dual-scoring prompt selection using LLMs and LALMs to enhance synthesis fidelity, semantic alignment, and speaker similarity.
• Experimental results show significant improvements in UTMOS, WER, and anti-spoofing metrics compared to baseline TTS systems.

Robust In-the-Wild TTS via EMA and Dual-Scoring Prompt Selection

Architectural Innovations

The paper introduces F5-TTS-DPS, a text-to-speech (TTS) system specifically optimized for in-the-wild data as part of the WildSpoof 2026 TTS Track. The backbone is F5-TTS (Chen et al., 2024), a large-scale, non-autoregressive model employing diffusion transformer and flow-matching for efficient high-fidelity speech synthesis. Adapting this architecture to noisy, real-world audio is achieved through two key enhancements: Exponential Moving Average (EMA) for training stabilization, and a dual-scoring prompt selection scheme employing LLMs and large audio LLMs (LALMs).

EMA, as implemented in this work, maintains a temporally smoothed average of parameters throughout fine-tuning, preventing overfitting and ensuring robust convergence despite inconsistencies in in-the-wild datasets [polyak1992acceleration] [tarvainen2017mean]. The dual-scoring prompt selection addresses prompt misalignment by evaluating both audio expressiveness (via Qwen2.5-Omni (Xu et al., 26 Mar 2025)) and semantic alignment (via Qwen3-30B-A3B (Yang et al., 14 May 2025)), filtering prompts to maximize synthesis fidelity and naturalness. This mechanism is crucial for voice cloning tasks, where prompt quality directly impacts generation consistency, especially in scenarios with degraded or misaligned recordings.

Experimental Methodology

Fine-tuning utilized the TITW-easy and TITW-hard subsets from the TITW dataset (Jung et al., 2024). Full-parameter training was performed, leveraging EMA (beta=0.99) and strategic hyperparameter selection (maximum batch size, 10 epochs, learning rate $1\times10^{-6}$, extensive warmup, and gradient clipping). Ablation studies systematically evaluated the contributions of SFT (supervised fine-tuning), EMA, and dual-scoring prompt selection. Metrics included UTMOS (Saeki et al., 2022), DNSMOS [reddy2021dnsmos], Word Error Rate (WER) via Whisper ASR (Radford et al., 2022), speaker similarity via ESPnet2 [jung2024espnet], SDS (anti-spoofing score via AASIST [wang2020asvspoof]), and a-DCF (Shim et al., 2024) for spoofing detectability.

Performance Evaluation

F5-TTS-DPS demonstrates significant improvements in synthesis quality, spoofing robustness, and intelligibility relative to both the F5-TTS baseline and leading challenge submissions. Key numerical outcomes include:

• UTMOS of 3.20 and WER of 8.65% on the development set, evidencing high naturalness and linguistic consistency.
• Speaker similarity improved to 0.508 (13% higher than baseline).
• SDS reduced to 0.108 (indicating substantial enhancement in authenticity and difficulty of detection).
• Best a-DCF scores (0.1582, 0.5233, 0.2562 across three SASV systems) among all submissions. This quantifies the model’s ability to generate speech with the highest spoofing potential while maintaining naturalness.

The ablation study further establishes that EMA fine-tuning and dual-scoring prompt selection are both indispensable; EMA contributes chiefly to stability and robustness, whereas prompt selection directly enhances output quality and semantic alignment. The combined effect results in a synthesized speech that is simultaneously natural, consistent, and resistant to state-of-the-art anti-spoofing systems.

Practical and Theoretical Implications

This work demonstrates that high-fidelity, robust TTS synthesis in unconstrained environments is achievable without reliance on meticulously curated studio training sets. EMA integration provides a scalable solution for parameter stabilization in large foundational models subjected to heterogeneous data distributions. The dual-scoring prompt selection, leveraging specialized LLM and LALM, positions prompt curation as a pivotal process for optimizing TTS in noisy, real-world conditions, and sets a precedent for future prompt engineering methodologies in voice cloning and expressive synthesis.

The strong a-DCF performance underscores both practical and security implications: generated speech is not only indistinguishable by humans but also by advanced SASV detectors, posing new challenges for anti-spoofing and voice authentication. The methods could enable broader TTS deployment in consumer devices, voice assistants, and accessibility tools, but will necessitate ongoing research in detection and countermeasures.

Future Directions

Advancing this framework will likely entail reinforcement learning to dynamically adapt to in-the-wild adversarial environments, promoting the co-evolution of synthesis and detection models. Further exploration of multi-modal prompt evaluation and adaptive selection heuristics could yield even more robust speech generation systems. The intertwining of TTS and SASV research suggests imminent opportunities for adversarial training paradigms and improved security architectures for voice-based authentication.

Conclusion

F5-TTS-DPS leverages EMA-stabilized fine-tuning and dual-scoring prompt selection to produce in-the-wild TTS that is both highly natural and challenging for spoofing detectors to distinguish. The integration of LLM and LALM techniques in the prompt selection process, alongside robust foundational modeling, sets a new technical baseline for speech synthesis under real-world conditions. The results call for concurrent advances in anti-spoofing research and prompt engineering, with substantial implications for the development and deployment of secure, realistic TTS technologies (2605.23859).